LLVM 19.1 will soon be released. This post provides a summary of my contributions in this release cycle to record my learning progress.
- LLVM MC: Integrated assembler improvements in LLVM 19
- lld/ELF: lld 19 ELF changes
- AArch64
- ARM
- RISC-V
- sanitizers
LLVM binary utilities
- [llvm-readobj,ELF] Support --decompress/-z
- [llvm-objcopy] Improve help messages
- [llvm-readelf] Print a blank line for the first hex/string dump
- [llvm-objcopy] Add --compress-sections
- [llvm-readelf] Print more information for RELR
Hashing
I optimized the bit mixer used by
llvm::DenseMap<std::pair<X, Y>> and
llvm::DenseMap<std::tuple<X...>>.
llvm/ADT/Hashing.h, used by StringRef hashing
and DenseMap, was supposed to be non-deterministic. Despite
this, a lot of code relied on a specific iteration order. I made
multiple fixes across the code base and landed [Hashing] Use a
non-deterministic seed if LLVM_ENABLE_ABI_BREAKING_CHECKS to improve
test coverage (e.g. assertion builds) and ensure future flexibility to
replace the algorithm.
The change has a noticeable code size reduction 1
2
3
4
5
6
7
8
9# old
movq _ZN4llvm7hashing6detail19fixed_seed_overrideE@GOTPCREL(%rip), %rax
movq (%rax), %rax
testq %rax, %rax
movabsq $-49064778989728563, %rcx # imm = 0xFF51AFD7ED558CCD
cmoveq %rcx, %rax
# new
movabsq $-49064778989728563, %rcx
... and significant compile time improvement.
I optimized
DenseMap::{find,erase}, yielding compile time
improvement.
Optimizations to the bit mixer in Hashing.h and the
DenseMap code have yielded significant benefits, reducing
both compile time and code size. This suggests there's further potential
for improvement in this area.
However, the reduced code size also highlights potential significant
code size increase when considering faster unordered map implementations
like boost::unordered_flat_map,
Abseil's Swiss
Table, and Folly's
F14. While these libraries may offer better performance, they often
come with a significant increase in code complexity and size.
Introducing a new container alongside DenseMap to
selectively replace performance-critical instances could lead to
substantial code modifications. This approach requires careful
consideration to balance potential performance gains with the additional
complexity.
NumericalStabilitySanitizer
NumericalStabilitySanitizer is a new feature for the 19.x releases. I have made many changes on the compiler-rt part.
Clang
Driver maintenance
- [Driver] Ensure ToolChain::LibraryPaths is not empty for non-Darwin
[Driver] Support -Wa,--defsym similar to -Wa,-defsym.- [Driver] Don't default to -mrelax-all for non-RISCV -O0. See Clang's -O0 output: branch displacement and size increase
Options used by the LLVM integrated assembler are currently handled
in an ad-hoc way. There is deduplication with and without LTO.
Eventually we might want to adopt TableGen for these -Wa,
options.
Others:
- [Sema] -Wpointer-bool-conversion: suppress lambda function pointer conversion diagnostic during instantiation
- [Sema] Mark alias/ifunc targets used and consider mangled names. This patch received a lot of scrutiny because it fixed an issue reported the XZ backdoor author(s).
- Make diagnostic pragma override -Werror=foo and DefaultError warnings
- Many patches migrated code off from potentially non-deterministic
llvm/ADT/Hashing.h - [CodeGen] Set attributes on resolvers emitted after ifuncs
Code review
I reviewed a wide range of patches, including areas like ADT/Support, binary utilities, MC, lld, clangDriver, LTO, sanitizers, LoongArch, RISC-V, and new features like NumericalStabilitySanitizer and RealTimeSanitizer.
To quantify my involvement, a search for patches I commented on
(repo:llvm/llvm-project is:pr -author:MaskRay commenter:MaskRay created:>2024-01-23)
yields 780 results.